2026-08-25

• 08:01 fetched 20260824-183902-qwen3_4b_gsm8k_fixed_batch_reward_signal_jean_zay-39c4380
• 21:36 fetched 20260824-183902-qwen3_4b_gsm8k_fixed_batch_reward_signal_jean_zay-39c4380
• 21:37 reported 20260824-183902-qwen3_4b_gsm8k_fixed_batch_reward_signal_jean_zay-39c4380 in notes/experiments/qwen3_4b_gsm8k_fixed_batch_reward_signal_jean_zay.md
• 21:37 fetched and reported the fixed-batch Qwen3 GSM8K reward-signal diagnostic; the positive learning-rate condition produced a clear center-model gain, while the negative learning-rate control failed at optimizer initialization